Situational Awareness LP (SALP), founded in 2024 with $225 million by former OpenAI researcher Leopold Aschenbrenner, peaked at $45 billion AUM in early July 2026 — then imploded. On July 30, 2026, the fund was forced to liquidate its entire public market portfolio in a single block trade to Citadel, losing 67% in one month. Yet the 165-page "investment bible" that launched Aschenbrenner's fame remains essential reading for anyone investing in AI. This report examines why the thesis was right, why the fund was wrong, and what the Cisco parallel teaches about hardware-first bets.
| Timeline | Event |
|---|---|
| 2024 | SALP founded with $225M |
| April 2024 | Aschenbrenner fired from OpenAI |
| June 2024 | Published 165-page "Situational Awareness: The Decade Ahead" |
| 2024–2026 | Fund returns >1,000% (epic run) |
| July 2026 | AUM peaks at $45 billion |
| July 30, 2026 | Entire public portfolio sold to Citadel in block trade |
| July 2026 | Monthly loss: 67% |
| Post-crash | Still holds ~$5 billion in Anthropic equity |
"The fall is relative." At $5 billion in remaining Anthropic equity, Aschenbrenner is hardly destitute. The "fall" is from mountaintop to mountainside.
In June 2024, Aschenbrenner published what Silicon Valley quickly dubbed the AI Investment Bible — a 165-page manifesto titled Situational Awareness: The Decade Ahead.
Key Predictions
| Prediction | Timeline | Status (2026) |
|---|---|---|
| AGI arrival | 2027 | Still pending |
| AI training cluster power consumption reaches 20% of U.S. generation | 2030 | On track |
| AI infrastructure investment scales to trillions | 2026–2030 | Partially validated |
| Self-improving AI research | Post-AGI | Not yet |
The Hardware-First Bet
Aschenbrenner put his money where his mouth was, constructing a portfolio that was long the entire AI infrastructure stack and short traditional software:
| Position | Examples | Thesis |
|---|---|---|
| Long: GPU clouds | CoreWeave, Nebius | Compute scarcity |
| Long: Memory/Storage | SanDisk, Micron, SK Hynix | Data infrastructure |
| Long: Power | Bloom Energy | Energy constraint |
| Long: Bitcoin miners → Data centers | IREN, Core Scientific | Power + land assets |
| Short: Legacy software | Adobe, etc. | AI disruption victims |
Aschenbrenner's fall echoes the Cisco bubble of 2000. During the dot-com era, Cisco was the "pick and shovel" play — the infrastructure provider for the internet revolution. Its stock peaked at $80, then crashed to $13. The internet thesis was correct; the timing and valuation were not.
Why Hardware-First Bets Fail
| Factor | Cisco 2000 | AI Infrastructure 2026 |
|---|---|---|
| Thesis | Internet will transform everything | AI will transform everything |
| Infrastructure built | Massive fiber, routers, switches | Massive data centers, GPUs |
| Revenue lag | Years of overcapacity | Emerging |
| Killer app gap | No revenue-generating apps yet | No AGI yet |
| Valuation | Priced for perfection | Priced for acceleration |
Infrastructure is the foundation of every technology revolution. But foundations are not returns.
The pattern is consistent: railroads in the 1870s, fiber optics in 2000, shale oil in 2014 — infrastructure buildouts create overcapacity before demand catches up. The winners are rarely the infrastructure builders; they are the applications that eventually run on top.
Every technology revolution needs its "killer app" — the entry point that makes the technology indispensable to ordinary users:
| Era | Killer App | What It Did |
|---|---|---|
| Electricity | Light bulb | Brought wires into every home |
| Internet | Browser | Made the web accessible to all |
| Mobile | iPhone | Put computing in every pocket |
| AI | ??? | Not yet clear |
Aschenbrenner argued that ChatGPT is not the killer app — impressive, but it has not fundamentally changed daily life. The author agrees, suggesting agentic coding (AI-assisted programming) as the closest current candidate — but this affects only programmers, not the general population.
The true grand prize is AGI itself: a system capable of any cognitive work, compressing the marginal cost of cognition to near zero.
Yann LeCun, Turing Award winner and former Meta Chief AI Scientist, offers a dissenting view. He argues that LLMs alone, relying solely on "predicting the next token," cannot generate true intelligence. His evidence: AI knows专业知识倒背如流 but lacks common sense, frequently hallucinating. OpenAI's own 2025 paper Why Language Models Hallucinate confirmed that hallucinations are inherent and uneradicable in LLMs.
LeCun's Solution: World Models
In late 2025, LeCun left Meta after 12 years to found AMI Labs, going all-in on "world models" — now valued at $3.5 billion.
| Capability | LLM | World Model |
|---|---|---|
| Knowledge | Extensive (text-based) | Moderate |
| Common sense | Poor | Designed for it |
| Physical reasoning | None | Core capability |
| Hallucination | Inherent | Minimized |
| Training data | Text | Multi-sensorial |
The Fusion Hypothesis
The author suggests the ultimate AGI architecture may be a fusion: LLMs handle language and knowledge; world models handle physical reality understanding. The world's smartest minds are currently working on this integration.
Aschenbrenner's investment bible will remain essential reading for AI investors. His structural analysis — the trillion-dollar infrastructure buildout, the energy constraints, the inevitability of AGI — is directionally correct.
But directionally correct does not mean immediately profitable. The Cisco lesson endures: being right about the revolution is not the same as being right about the timing, the valuations, or the specific winners.
For investors today, the takeaway is: read the bible for the map, but navigate with a valuation compass.
Standard Kepler Research | standardkepler.com